Design of an optimal Chebyshev-expanded discrimination function for globular proteins

Optimizing energy potential for protein fold recognition with parametric evaluation function

In this paper, a new optimization method is proposed to determine a simplified energy potential for protein fold recognition, which consists of the residue-residue contact, hydrophobicity, and pseudodihedral potentials. With a parametric evaluation function method, the Z-scores of all the proteins in a training set are optimized simultaneously to obtain the best parameter set of the potential. For this multi-objective and multi-constraint problem, the new optimization scheme is very effective. The derived potential is then tested on two high-quality decoy sets and compared with other classical fold recognition potentials. With the simplified energy potential, we achieve a high level of discrimination capability between correct and incorrect folds.

Validation of protein models by a neural network approach

Background

The development and improvement of reliable computational methods designed to evaluate the quality of protein models is relevant in the context of protein structure refinement, which has been recently identified as one of the bottlenecks limiting the quality and usefulness of protein structure prediction.

Results

In this contribution, we present a computational method (Artificial Intelligence Decoys Evaluator: AIDE) which is able to consistently discriminate between correct and incorrect protein models. In particular, the method is based on neural networks that use as input 15 structural parameters, which include energy, solvent accessible surface, hydrophobic contacts and secondary structure content. The results obtained with AIDE on a set of decoy structures were evaluated using statistical indicators such as Pearson correlation coefficients, Z_nat, fraction enrichment, as well as ROC plots. It turned out that AIDE performances are comparable and often complementary to available state-of-the-art learning-based methods.

Conclusion

In light of the results obtained with AIDE, as well as its comparison with available learning-based methods, it can be concluded that AIDE can be successfully used to evaluate the quality of protein structures. The use of AIDE in combination with other evaluation tools is expected to further enhance protein refinement efforts.

How effective for fold recognition is a potential of mean force that includes relative orientations between contacting residues in proteins?

We estimate the statistical distribution of relative orientations between contacting residues from a database of protein structures and evaluate the potential of mean force for relative orientations between contacting residues. Polar angles and Euler angles are used to specify two degrees of directional freedom and three degrees of rotational freedom for the orientation of one residue relative to another in contacting residues, respectively. A local coordinate system affixed to each residue based only on main chain atoms is defined for fold recognition. The number of contacting residue pairs in the database will severely limit the resolution of the statistical distribution of relative orientations, if it is estimated by dividing space into cells and counting samples observed in each cell. To overcome such problems and to evaluate the fully anisotropic distributions of relative orientations as a function of polar and Euler angles, we choose a method in which the observed distribution is represented as a sum of delta functions each of which represents the observed orientation of a contacting residue, and is evaluated as a series expansion of spherical harmonics functions. The sample size limits the frequencies of modes whose expansion coefficients can be reliably estimated. High frequency modes are statistically less reliable than low frequency modes. Each expansion coefficient is separately corrected for the sample size according to suggestions from a Bayesian statistical analysis. As a result, many expansion terms can be utilized to evaluate orientational distributions. Also, unlike other orientational potentials, the uniform distribution is used for a reference distribution in evaluating a potential of mean force for each type of contacting residue pair from its orientational distribution, so that residue-residue orientations can be fully evaluated. It is shown by using decoy sets that the discrimination power of the orientational potential in fold recognition increases by taking account of the Euler angle dependencies and becomes comparable to that of a simple contact potential, and that the total energy potential taken as a simple sum of contact, orientation, and (phi,psi) potentials performs well to identify the native folds.

Monte Carlo sampling of near-native structures of proteins with applications

Since a protein's dynamic fluctuation inside cells affects the protein's biological properties, we present a novel method to study the ensemble of near-native structures (NNS) of proteins, namely, the conformations that are very similar to the experimentally determined native structure. We show that this method enables us to (i) quantify the difficulty of predicting a protein's structure, (ii) choose appropriate simplified representations of protein structures, and (iii) assess the effectiveness of knowledge-based potential functions. We found that well-designed simple representations of protein structures are likely as accurate as those more complex ones for certain potential functions. We also found that the widely used contact potential functions stabilize NNS poorly, whereas potential functions incorporating local structure information significantly increase the stability of NNS.

Empirical potential function for simplified protein models: combining contact and local sequence-structure descriptors

An effective potential function is critical for protein structure prediction and folding simulation. Simplified protein models such as those requiring only Calpha or backbone atoms are attractive because they enable efficient search of the conformational space. We show residue-specific reduced discrete-state models can represent the backbone conformations of proteins with small RMSD values. However, no potential functions exist that are designed for such simplified protein models. In this study, we develop optimal potential functions by combining contact interaction descriptors and local sequence-structure descriptors. The form of the potential function is a weighted linear sum of all descriptors, and the optimal weight coefficients are obtained through optimization using both native and decoy structures. The performance of the potential function in a test of discriminating native protein structures from decoys is evaluated using several benchmark decoy sets. Our potential function requiring only backbone atoms or Calpha atoms have comparable or better performance than several residue-based potential functions that require additional coordinates of side-chain centers or coordinates of all side-chain atoms. By reducing the residue alphabets down to size 10 for contact descriptors, the performance of the potential function can be further improved. Our results also suggest that local sequence-structure correlation may play important role in reducing the entropic cost of protein folding.

Geometric cooperativity and anticooperativity of three-body interactions in native proteins

Characterizing multibody interactions of hydrophobic, polar, and ionizable residues in protein is important for understanding the stability of protein structures. We introduce a geometric model for quantifying 3-body interactions in native proteins. With this model, empirical propensity values for many types of 3-body interactions can be reliably estimated from a database of native protein structures, despite the overwhelming presence of pairwise contacts. In addition, we define a nonadditive coefficient that characterizes cooperativity and anticooperativity of residue interactions in native proteins by measuring the deviation of 3-body interactions from 3 independent pairwise interactions. It compares the 3-body propensity value from what would be expected if only pairwise interactions were considered, and highlights the distinction of propensity and cooperativity of 3-body interaction. Based on the geometric model, and what can be inferred from statistical analysis of such a model, we find that hydrophobic interactions and hydrogen-bonding interactions make nonadditive contributions to protein stability, but the nonadditive nature depends on whether such interactions are located in the protein interior or on the protein surface. When located in the interior, many hydrophobic interactions such as those involving alkyl residues are anticooperative. Salt-bridge and regular hydrogen-bonding interactions, such as those involving ionizable residues and polar residues, are cooperative. When located on the protein surface, these salt-bridge and regular hydrogen-bonding interactions are anticooperative, and hydrophobic interactions involving alkyl residues become cooperative. We show with examples that incorporating 3-body interactions improves discrimination of protein native structures against decoy conformations. In addition, analysis of cooperative 3-body interaction may reveal spatial motifs that can suggest specific protein functions.

Improved protein structure selection using decoy-dependent discriminatory functions

BACKGROUND

A key component in protein structure prediction is a scoring or discriminatory function that can distinguish near-native conformations from misfolded ones. Various types of scoring functions have been developed to accomplish this goal, but their performance is not adequate to solve the structure selection problem. In addition, there is poor correlation between the scores and the accuracy of the generated conformations.

RESULTS

We present a simple and nonparametric formula to estimate the accuracy of predicted conformations (or decoys). This scoring function, called the density score function, evaluates decoy conformations by performing an all-against-all Calpha RMSD (Root Mean Square Deviation) calculation in a given decoy set. We tested the density score function on 83 decoy sets grouped by their generation methods (4state_reduced, fisa, fisa_casp3, lmds, lattice_ssfit, semfold and Rosetta). The density scores have correlations as high as 0.9 with the Calpha RMSDs of the decoy conformations, measured relative to the experimental conformation for each decoy. We previously developed a residue-specific all-atom probability discriminatory function (RAPDF), which compiles statistics from a database of experimentally determined conformations, to aid in structure selection. Here, we present a decoy-dependent discriminatory function called self-RAPDF, where we compiled the atom-atom contact probabilities from all the conformations in a decoy set instead of using an ensemble of native conformations, with a weighting scheme based on the density scores. The self-RAPDF has a higher correlation with Calpha RMSD than RAPDF for 76/83 decoy sets, and selects better near-native conformations for 62/83 decoy sets. Self-RAPDF may be useful not only for selecting near-native conformations from decoy sets, but also for fold simulations and protein structure refinement.

CONCLUSIONS

Both the density score and the self-RAPDF functions are decoy-dependent scoring functions for improved protein structure selection. Their success indicates that information from the ensemble of decoy conformations can be used to derive statistical probabilities and facilitate the identification of near-native structures.

Mimicking the action of folding chaperones in molecular dynamics simulations: Application to the refinement of homology-based protein structures

A novel method for the refinement of misfolded protein structures is proposed in which the properties of the solvent environment are oscillated in order to mimic some aspects of the role of molecular chaperones play in protein folding in vivo. Specifically, the hydrophobicity of the solvent is cycled by repetitively altering the partial charges on solvent molecules (water) during a molecular dynamics simulation. During periods when the hydrophobicity of the solvent is increased, intramolecular hydrogen bonding and secondary structure formation are promoted. During periods of increased solvent polarity, poorly packed regions of secondary structures are destabilized, promoting structural rearrangement. By cycling between these two extremes, the aim is to minimize the formation of long-lived intermediates. The approach has been applied to the refinement of structural models of three proteins generated by using the ROSETTA procedure for ab initio structure prediction. A significant improvement in the deviation of the model structures from the corresponding experimental structures was observed. Although preliminary, the results indicate computationally mimicking some functions of molecular chaperones in molecular dynamics simulations can promote the correct formation of secondary structure and thus be of general use in protein folding simulations and in the refinement of structural models of small- to medium-size proteins.

Optimizing physical energy functions for protein folding

We optimize a physical energy function for proteins with the use of the available structural database and perform three benchmark tests of the performance: (1) recognition of native structures in the background of predefined decoy sets of Levitt, (2) de novo structure prediction using fragment assembly sampling, and (3) molecular dynamics simulations. The energy parameter optimization is based on the energy landscape theory and uses a Monte Carlo search to find a set of parameters that seeks the largest ratio deltaE(s)/DeltaE for all proteins in a training set simultaneously. Here, deltaE(s) is the stability gap between the native and the average in the denatured states and DeltaE is the energy fluctuation among these states. Some of the energy parameters optimized are found to show significant correlation with experimentally observed quantities: (1) In the recognition test, the optimized function assigns the lowest energy to either the native or a near-native structure among many decoy structures for all the proteins studied. (2) Structure prediction with the fragment assembly sampling gives structure models with root mean square deviation less than 6 A in one of the top five cluster centers for five of six proteins studied. (3) Structure prediction using molecular dynamics simulation gives poorer performance, implying the importance of having a more precise description of local structures. The physical energy function solely inferred from a structural database neither utilizes sequence information from the family of the target nor the outcome of the secondary structure prediction but can produce the correct native fold for many small proteins.

Funnel sculpting for in silico assembly of secondary structure elements of proteins

We present a method for designing a funnel-shaped free-energy surface that reproducibly assembles secondary structure elements of proteins into their native conformations from a random extended configuration. Assuming a priori knowledge of secondary structure, our method can design a funnel-shaped surface for folding of alpha, beta, and alphabeta structures individually. We design energy surfaces that fold up to five unrelated sequences with the same energy parameters. We develop a measure of the foldability of an energy landscape in silico and present an alternative way to view energy landscapes.

A simple procedure to weight empirical potentials in a fitness function so as to optimize its performance in ab initio protein-folding problem

In an approach to the protein folding problem by a Genetic Algorithm, the fitness function plays a critical role. Empirical potentials are generally used to build the fitness function, and they must be weighted to obtain a valuable one. The weights are generally found by the comparison with a set of misfolded structures (decoys), but a dependence of the obtained fitness generally arises on the used decoys. Here we describe a general procedure to find out, from a given set of potentials, their better linear combination that could either identify the wild structure or prove their powerlessness. We use topological considerations over the hyperspace of the potentials, and a multiple linear inequalities solver. The iterated method flows through the following steps: it determines a direction in the hyperspace of the potentials, which identifies the native structure as a vertex among a set of misfolded decoys. A multiple linear inequalities solver obtains the direction. A Genetic Algorithm, tailored to the specific problem, uses the fitness function defined by this direction and generally reaches a new structure better than the experimental one, which is added to the ensemble. The decoys so generated are not dependent on a deterministic criterion. This iterative procedure can be stopped either by identifying an effective fitness function or by proving the impossibility of its achievement. In order to test the method under the hardest conditions, we choose numerous and heterogeneous quantities as components of the fitness function. This method could be a useful tool for the scientific community in order to test any fitness proposed and to recognize the most important components on which it is built.

Can correct protein models be identified?

The ability to separate correct models of protein structures from less correct models is of the greatest importance for protein structure prediction methods. Several studies have examined the ability of different types of energy function to detect the native, or native-like, protein structure from a large set of decoys. In contrast to earlier studies, we examine here the ability to detect models that only show limited structural similarity to the native structure. These correct models are defined by the existence of a fragment that shows significant similarity between this model and the native structure. It has been shown that the existence of such fragments is useful for comparing the performance between different fold recognition methods and that this performance correlates well with performance in fold recognition. We have developed ProQ, a neural-network-based method to predict the quality of a protein model that extracts structural features, such as frequency of atom-atom contacts, and predicts the quality of a model, as measured either by LGscore or MaxSub. We show that ProQ performs at least as well as other measures when identifying the native structure and is better at the detection of correct models. This performance is maintained over several different test sets. ProQ can also be combined with the Pcons fold recognition predictor (Pmodeller) to increase its performance, with the main advantage being the elimination of a few high-scoring incorrect models. Pmodeller was successful in CASP5 and results from the latest LiveBench, LiveBench-6, indicating that Pmodeller has a higher specificity than Pcons alone.

Peptide backbone reconstruction using dead-end elimination and a knowledge-based forcefield

A novel, yet simple and automated, protocol for reconstruction of complete peptide backbones from C(alpha) coordinates only is described, validated, and benchmarked. The described method collates a set of possible backbone conformations for each set of residue triads from a structural library derived from the PDB. The optimal permutation of these three residue segments of backbone conformations is determined using the dead-end elimination (DEE) algorithm. Putative conformations are evaluated using a pairwise-additive knowledge-based forcefield term and a fragment overlap term. The protocol described in this report is able to restore the full backbone coordinates to within 0.2-0.6 A of the actual crystal structure from C(alpha) coordinates only. In addition, it is insensitive to errors in the input C(alpha) coordinates with RMSDs of 3.0 A, and this is illustrated through application to deliberately distorted C(alpha) traces. The entire process, as described, is rapid, requiring of the order of a few minutes for a typical protein on a typical desktop PC. Approximations enable this to be reduced to a few seconds, although this is at the expense of prediction accuracy. This compares very favorably to previously published methods, being sufficiently fast for general use and being one of the most accurate methods. Because the method is not restricted to the reconstruction from only C(alpha) coordinates, reconstruction based on C(beta) coordinates is also demonstrated.